[AArch64] Refactor the NEON scalar floating-point reciprocal estimate, floating-

point reciprocal exponent, and floating-point reciprocal square root estimate
LLVM AArch64 intrinsics to use f32/f64 types, rather than their vector
equivalents.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197066 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/include/llvm/IR/IntrinsicsAArch64.td b/include/llvm/IR/IntrinsicsAArch64.td
index 37abbe7db8ae5c651ac7d83323928d2fca79f02e..fb31452fb2392c83b86ae805fd1cfe67cc82115a 100644 (file)
--- a/include/llvm/IR/IntrinsicsAArch64.td
+++ b/include/llvm/IR/IntrinsicsAArch64.td
@@ -260,8 +260,17 @@ def int_aarch64_neon_fcvtzs :
 def int_aarch64_neon_fcvtzu :
   Intrinsic<[llvm_anyvector_ty], [llvm_anyfloat_ty], [IntrNoMem]>;
 
+// Scalar Floating-point Reciprocal Estimate.
+def int_aarch64_neon_vrecpe :
+  Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+
 // Scalar Floating-point Reciprocal Exponent
-def int_aarch64_neon_vrecpx : Neon_1Arg_Intrinsic;
+def int_aarch64_neon_vrecpx :
+  Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+
+// Scalar Floating-point Reciprocal Square Root Estimate
+def int_aarch64_neon_vrsqrte :
+  Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 
 class Neon_Cmp_Intrinsic
   : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_anyvector_ty],

diff --git a/lib/Target/AArch64/AArch64InstrNEON.td b/lib/Target/AArch64/AArch64InstrNEON.td
index 8f25818226c750fe73c44c270fac0f17a618d08a..9e02dc4fff179803d2c4e385048fa768b9865080 100644 (file)
--- a/lib/Target/AArch64/AArch64InstrNEON.td
+++ b/lib/Target/AArch64/AArch64InstrNEON.td
@@ -4346,12 +4346,17 @@ multiclass Neon_Scalar2SameMisc_cvt_SD_size_patterns<SDPatternOperator opnode,
 multiclass Neon_Scalar2SameMisc_SD_size_patterns<SDPatternOperator opnode,
                                                  Instruction INSTS,
                                                  Instruction INSTD> {
-  def : Pat<(v1f32 (opnode (v1f32 FPR32:$Rn))),
+  def : Pat<(f32 (opnode (f32 FPR32:$Rn))),
             (INSTS FPR32:$Rn)>;
-  def : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn))),
+  def : Pat<(f64 (opnode (f64 FPR64:$Rn))),
             (INSTD FPR64:$Rn)>;
 }
 
+class Neon_Scalar2SameMisc_V1_D_size_patterns<SDPatternOperator opnode,
+                                              Instruction INSTD>
+  : Pat<(v1f64 (opnode (v1f64 FPR64:$Rn))),
+        (INSTD FPR64:$Rn)>;
+
 class NeonI_Scalar2SameMisc_cmpz_D_size<bit u, bits<5> opcode, string asmop>
   : NeonI_Scalar2SameMisc<u, 0b11, opcode,
                           (outs FPR64:$Rd), (ins FPR64:$Rn, neon_uimm0:$Imm),
@@ -4592,8 +4597,8 @@ multiclass Neon_ScalarShiftLImm_D_size_patterns<SDPatternOperator opnode,
                 (INSTD FPR64:$Rn, imm:$Imm)>;
 }
 
-class Neon_ScalarShiftImm_arm_D_size_patterns<SDPatternOperator opnode,
-                                              Instruction INSTD>
+class Neon_ScalarShiftImm_V1_D_size_patterns<SDPatternOperator opnode,
+                                             Instruction INSTD>
   : Pat<(v1i64 (opnode (v1i64 FPR64:$Rn),
             (v1i64 (Neon_vdup (i32 shr_imm64:$Imm))))),
         (INSTD FPR64:$Rn, imm:$Imm)>;
@@ -4659,13 +4664,13 @@ multiclass Neon_ScalarShiftImm_fcvts_SD_size_patterns<SDPatternOperator opnode,
 defm SSHR : NeonI_ScalarShiftRightImm_D_size<0b0, 0b00000, "sshr">;
 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vshrds_n, SSHRddi>;
 // Pattern to match llvm.arm.* intrinsic.
-def : Neon_ScalarShiftImm_arm_D_size_patterns<sra, SSHRddi>;
+def : Neon_ScalarShiftImm_V1_D_size_patterns<sra, SSHRddi>;
 
 // Scalar Unsigned Shift Right (Immediate)
 defm USHR : NeonI_ScalarShiftRightImm_D_size<0b1, 0b00000, "ushr">;
 defm : Neon_ScalarShiftRImm_D_size_patterns<int_aarch64_neon_vshrdu_n, USHRddi>;
 // Pattern to match llvm.arm.* intrinsic.
-def : Neon_ScalarShiftImm_arm_D_size_patterns<srl, USHRddi>;
+def : Neon_ScalarShiftImm_V1_D_size_patterns<srl, USHRddi>;
 
 // Scalar Signed Rounding Shift Right (Immediate)
 defm SRSHR : NeonI_ScalarShiftRightImm_D_size<0b0, 0b00100, "srshr">;
@@ -4699,7 +4704,7 @@ def : Neon_ScalarShiftRImm_accum_D_size_patterns
 defm SHL : NeonI_ScalarShiftLeftImm_D_size<0b0, 0b01010, "shl">;
 defm : Neon_ScalarShiftLImm_D_size_patterns<int_aarch64_neon_vshld_n, SHLddi>;
 // Pattern to match llvm.arm.* intrinsic.
-def : Neon_ScalarShiftImm_arm_D_size_patterns<shl, SHLddi>;
+def : Neon_ScalarShiftImm_V1_D_size_patterns<shl, SHLddi>;
 
 // Signed Saturating Shift Left (Immediate)
 defm SQSHL : NeonI_ScalarShiftLeftImm_BHSD_size<0b0, 0b01110, "sqshl">;
@@ -5056,8 +5061,10 @@ def : Neon_Scalar2SameMisc_fcvt_v1f64_pattern<fp_to_uint, FCVTZUdd>;
 
 // Scalar Floating-point Reciprocal Estimate
 defm FRECPE : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11101, "frecpe">;
-defm : Neon_Scalar2SameMisc_SD_size_patterns<int_arm_neon_vrecpe,
+defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrecpe,
                                              FRECPEss, FRECPEdd>;
+def : Neon_Scalar2SameMisc_V1_D_size_patterns<int_arm_neon_vrecpe,
+                                              FRECPEdd>;
 
 // Scalar Floating-point Reciprocal Exponent
 defm FRECPX : NeonI_Scalar2SameMisc_SD_size<0b0, 0b1, 0b11111, "frecpx">;
@@ -5066,8 +5073,10 @@ defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrecpx,
 
 // Scalar Floating-point Reciprocal Square Root Estimate
 defm FRSQRTE: NeonI_Scalar2SameMisc_SD_size<0b1, 0b1, 0b11101, "frsqrte">;
-defm : Neon_Scalar2SameMisc_SD_size_patterns<int_arm_neon_vrsqrte,
-                                             FRSQRTEss, FRSQRTEdd>;
+defm : Neon_Scalar2SameMisc_SD_size_patterns<int_aarch64_neon_vrsqrte,
+                                                 FRSQRTEss, FRSQRTEdd>;
+def : Neon_Scalar2SameMisc_V1_D_size_patterns<int_arm_neon_vrsqrte,
+                                              FRSQRTEdd>;
 
 // Scalar Floating-point Round
 class Neon_ScalarFloatRound_pattern<SDPatternOperator opnode, Instruction INST>

diff --git a/test/CodeGen/AArch64/neon-scalar-recip.ll b/test/CodeGen/AArch64/neon-scalar-recip.ll
index f21c27bee435df2b87989b839e2b04b875bad946..bd549a86a40c164d704b09ebebdd0bf366fdd1fd 100644 (file)
--- a/test/CodeGen/AArch64/neon-scalar-recip.ll
+++ b/test/CodeGen/AArch64/neon-scalar-recip.ll
@@ -50,9 +50,7 @@ define float @test_vrecpes_f32(float %a) {
 ; CHECK: test_vrecpes_f32
 ; CHECK: frecpe {{s[0-9]+}}, {{s[0-9]+}}
 entry:
-  %vrecpe.i = insertelement <1 x float> undef, float %a, i32 0
-  %vrecpe1.i = tail call <1 x float> @llvm.arm.neon.vrecpe.v1f32(<1 x float> %vrecpe.i)
-  %0 = extractelement <1 x float> %vrecpe1.i, i32 0
+  %0 = call float @llvm.aarch64.neon.vrecpe.f32(float %a)
   ret float %0
 }
 
@@ -60,22 +58,18 @@ define double @test_vrecped_f64(double %a) {
 ; CHECK: test_vrecped_f64
 ; CHECK: frecpe {{d[0-9]+}}, {{d[0-9]+}}
 entry:
-  %vrecpe.i = insertelement <1 x double> undef, double %a, i32 0
-  %vrecpe1.i = tail call <1 x double> @llvm.arm.neon.vrecpe.v1f64(<1 x double> %vrecpe.i)
-  %0 = extractelement <1 x double> %vrecpe1.i, i32 0
+  %0 = call double @llvm.aarch64.neon.vrecpe.f64(double %a)
   ret double %0
 }
 
-declare <1 x float> @llvm.arm.neon.vrecpe.v1f32(<1 x float>)
-declare <1 x double> @llvm.arm.neon.vrecpe.v1f64(<1 x double>)
+declare float @llvm.aarch64.neon.vrecpe.f32(float)
+declare double @llvm.aarch64.neon.vrecpe.f64(double)
 
 define float @test_vrecpxs_f32(float %a) {
 ; CHECK: test_vrecpxs_f32
 ; CHECK: frecpx {{s[0-9]+}}, {{s[0-9]+}}
 entry:
-  %vrecpx.i = insertelement <1 x float> undef, float %a, i32 0
-  %vrecpx1.i = tail call <1 x float> @llvm.aarch64.neon.vrecpx.v1f32(<1 x float> %vrecpx.i)
-  %0 = extractelement <1 x float> %vrecpx1.i, i32 0
+  %0 = call float @llvm.aarch64.neon.vrecpx.f32(float %a)
   ret float %0
 }
 
@@ -83,22 +77,18 @@ define double @test_vrecpxd_f64(double %a) {
 ; CHECK: test_vrecpxd_f64
 ; CHECK: frecpx {{d[0-9]+}}, {{d[0-9]+}}
 entry:
-  %vrecpx.i = insertelement <1 x double> undef, double %a, i32 0
-  %vrecpx1.i = tail call <1 x double> @llvm.aarch64.neon.vrecpx.v1f64(<1 x double> %vrecpx.i)
-  %0 = extractelement <1 x double> %vrecpx1.i, i32 0
+  %0 = call double @llvm.aarch64.neon.vrecpx.f64(double %a)
   ret double %0
 }
 
-declare <1 x float> @llvm.aarch64.neon.vrecpx.v1f32(<1 x float>)
-declare <1 x double> @llvm.aarch64.neon.vrecpx.v1f64(<1 x double>)
+declare float @llvm.aarch64.neon.vrecpx.f32(float)
+declare double @llvm.aarch64.neon.vrecpx.f64(double)
 
 define float @test_vrsqrtes_f32(float %a) {
 ; CHECK: test_vrsqrtes_f32
 ; CHECK: frsqrte {{s[0-9]+}}, {{s[0-9]+}}
 entry:
-  %vrsqrte.i = insertelement <1 x float> undef, float %a, i32 0
-  %vrsqrte1.i = tail call <1 x float> @llvm.arm.neon.vrsqrte.v1f32(<1 x float> %vrsqrte.i)
-  %0 = extractelement <1 x float> %vrsqrte1.i, i32 0
+  %0 = call float @llvm.aarch64.neon.vrsqrte.f32(float %a)
   ret float %0
 }
 
@@ -106,11 +96,9 @@ define double @test_vrsqrted_f64(double %a) {
 ; CHECK: test_vrsqrted_f64
 ; CHECK: frsqrte {{d[0-9]+}}, {{d[0-9]+}}
 entry:
-  %vrsqrte.i = insertelement <1 x double> undef, double %a, i32 0
-  %vrsqrte1.i = tail call <1 x double> @llvm.arm.neon.vrsqrte.v1f64(<1 x double> %vrsqrte.i)
-  %0 = extractelement <1 x double> %vrsqrte1.i, i32 0
+  %0 = call double @llvm.aarch64.neon.vrsqrte.f64(double %a)
   ret double %0
 }
 
-declare <1 x float> @llvm.arm.neon.vrsqrte.v1f32(<1 x float>)
-declare <1 x double> @llvm.arm.neon.vrsqrte.v1f64(<1 x double>)
+declare float @llvm.aarch64.neon.vrsqrte.f32(float)
+declare double @llvm.aarch64.neon.vrsqrte.f64(double)